26 Nov 2014 – IGI Journal Club
Automatic Deformable MR-Ultrasound Registration
for Image-Guided Neurosurgery
By H. Rivaz, S. Chen and D. L. Collins in TMI 2013
Presented by: Rachel Sparks
26 Nov 2014 – IGI Journal Club
Outline
1. Motivation for MRI-US
2. RaPTOR
a) Correlation Ratio (CR)
b) CR Extension to Patches
c) Outlier Suppression
3. Experimental Design & Results
4. Concluding Remarks
26 Nov 2014 – IGI Journal Club
Clinical Motivation:
• Improved patient outcome with
complete tumor removal
Poor visibility of tumor boundary
Tumor proximity to critical brain
structures (vessels, eloquent
regions)
• Pre-operative MRI planning:
Location of the tumor boundary
Location of critical structures
• How to provide intra-operative
feedback and?
26 Nov 2014 – IGI Journal Club
Intraoperative imaging approaches
• Intraoperative MRI
High quality, visually similar to pre-operative MRI
Expensive, disruptive to surgery
• Intraoperative video/LRS
Restricted to the superficial surfaces
Cheap, non-disruptive
• Intraoperative US
Intensity inhomogeneity (i.e. signal attenuation, shadowing)
Difficult to interpret (speckle, low-resolution, limited FOV)
Provides minimally disruptive volumetric information
26 Nov 2014 – IGI Journal Club
26 Nov 2014 – IGI Journal Club
Previous Work MRI-US Fusion
• Surface-based registration (Reinertsen 2007):
Segment structures of interest (vessels) on both modalities
Register surfaces (FEM, ICP)
• Pseduo-US registration (Arbel 2004, Mercier 2012) :
From MRI general pseudo-US
Register pseudo-US to US (ANIMAL, block matching)
• Multimodal image registration
LMI (Klein 2007), Self-Similarity (Rivaz 2012), CR (Roche 1998)
Feature extraction and matching (Penney 2004, Coupé 2012)
26 Nov 2014 – IGI Journal Club
Novel Contributions
• Robust Patch-based Correlation Ratio (RaPTOR)- a local
correlation ratio metric
• Derive an analytic derivative for efficient optimization
• Perform outlier rejection to improve robustness of RaPTOR
26 Nov 2014 – IGI Journal Club
General Registration Formulation
• Transformation(𝑇) that minimizes dissimilarity metric (𝐷) and
regularization constraint (𝑅)
• In this work 𝑇 is modelled as a Free-form Deformation (FFD)
• Novelty is RaPTOR as choice of 𝐷
𝑇 = min𝑇𝐷 𝐼𝑓 𝑥 , 𝐼𝑚 𝑇 𝜙, 𝑥 + 𝑅 𝑇𝑇 = min
𝑇𝐷 𝐼𝑓 𝑥 , 𝐼𝑚 𝑇 𝜙, 𝑥 +
𝜔𝑅2𝑡𝑟𝑎𝑐𝑒 𝛻𝑇𝑇𝛻𝑇
26 Nov 2014 – IGI Journal Club
Correlation Ratio (CR)
• 𝑋, 𝑌 pixel intensities in 𝐼𝑓 and 𝐼𝑚 (choice of correspondence)
CR provides a value of what information in 𝑌 is explained by 𝑋
Define E 𝑌|𝑋 as expected value of 𝑌 given 𝑋*
• CR is calculated as,
𝜂 𝑌 𝑋 =Var[E 𝑌|𝑋 ]
Var[𝑌]
*Different techniques to estimate E 𝑌|𝑋
= 1 −Var[𝑌 − E 𝑌|𝑋 ]
Var[𝑌]
26 Nov 2014 – IGI Journal Club
Estimating E 𝑌|𝑋
• A non-parametric approach chosen – flexible, does not assume
specific intensity relationships
• Bin 𝑋 values to estimate 𝑌 value (Parzen windowing)
𝑋 𝑌
26 Nov 2014 – IGI Journal Club
Estimating E 𝑌|𝑋
• Binning allows closed form solution of 𝜂 𝑌 𝑋 ,
1 − 𝜂 𝑌 𝑋 =1
𝑁 𝜎2
𝑖=1
𝑁
𝑦𝑖2 −
𝑗=1
𝑁𝑏
𝑁𝑗𝜇𝑗
• 𝜎2 variance of 𝑌
26 Nov 2014 – IGI Journal Club
Estimating E 𝑌|𝑋
• Binning allows closed form solution of 𝜂 𝑌 𝑋 ,
1 − 𝜂 𝑌 𝑋 =1
𝑁 𝜎2
𝑖=1
𝑁
𝑦𝑖2 −
𝑗=1
𝑁𝑏
𝑁𝑗𝜇𝑗
• average of 𝑌
26 Nov 2014 – IGI Journal Club
Estimating E 𝑌|𝑋
• Binning allows closed form solution of 𝜂 𝑌 𝑋 ,
1 − 𝜂 𝑌 𝑋 =1
𝑁 𝜎2
𝑖=1
𝑁
𝑦𝑖2 −
𝑗=1
𝑁𝑏
𝑁𝑗𝜇𝑗
• estimated average of 𝑌 according to 𝑋 values (binning)
26 Nov 2014 – IGI Journal Club
Patch-based Correlation Ratio
• CR assumes pixel correlations consistent across the image
• US intensities vary in a spatially dependent manner
• Calculate CR independently for several small patches selected
through out the volume
RaPTOR 𝑋, 𝑌 =1
𝑁𝑝
𝑖=1
𝑁𝑝
(1 − 𝜂 𝑌 𝑋: Ω𝑖)
26 Nov 2014 – IGI Journal Club
Optimization of 𝐷
• To allow for efficient optimization (gradient descent) need to
define derivative of 𝐷 (chain rule),
𝜕𝐷
𝜕𝑇=𝜕𝑇
𝜕𝜙⋅𝜕𝐼𝑚𝜕𝑇⋅𝜕𝐷
𝜕𝐼𝑚
•𝜕𝐷
𝜕𝐼𝑚needs to be calculated from the RaPTOR metric
26 Nov 2014 – IGI Journal Club
Derivative of 𝐷
• Dependent on if 𝑋 corresponds to 𝐼𝑓 or 𝐼𝑚
• In this work 𝑋 corresponds to 𝐼𝑓 (MRI) so,
𝜕𝐷
𝜕𝐼𝑚=𝜕 1 − 𝜂 𝑌 𝑋
𝜕𝑦𝑖
• Once again using the chain rule
𝜕𝐷
𝜕𝐼𝑚=−𝜕𝜎2
𝜕𝑦𝑖
𝑁 𝜎4 𝑖=1𝑁 𝑦𝑖
2 − 𝑗=1𝑁𝑏 𝑁𝑗𝜇𝑗 +
1
𝑁 𝜎2
𝜕 𝑖=1𝑁 𝑦𝑖
2 − 𝑗=1
𝑁𝑏 𝑁𝑗𝜇𝑗
𝜕𝑦𝑖
26 Nov 2014 – IGI Journal Club
Robust Patch-based Correlation Ratio
• Some patches provide a poor estimate of correspondence
• Reject outlier patches to improve robustness of CR
• CR is a poor outlier detector
CR varies according to image alignment
Low CR could be misalignment (important to include) or mismatched
patch (outlier to ignore)
• Novel outlier detection according to CR gradient
26 Nov 2014 – IGI Journal Club
Robust Patch-based Correlation Ratio
• CR descent gradient (how image transformation is updated)
should be consistent between neighbouring pixels
26 Nov 2014 – IGI Journal Club
Robust Patch-based Correlation Ratio
• CR descent gradient (how image transformation is updated)
should be consistent between neighbouring pixels
26 Nov 2014 – IGI Journal Club
Robust Patch-based Correlation Ratio
• Formal the gradient descent direction is,𝛻𝐷
𝛻𝑇=𝛻𝐼𝑚𝛻𝑇⋅𝛻𝐷
𝛻𝐼𝑚• This is converted in to a unitless metric
𝑟 = minVar𝜕𝐷𝜕𝑇𝑥
𝜕𝐷𝜕𝑇𝑥
2 ,Var𝜕𝐷𝜕𝑇𝑌
𝜕𝐷𝜕𝑇𝑌
2 ,Var𝜕𝐷𝜕𝑇𝑧
𝜕𝐷𝜕𝑇𝑧
2
• Patches are rejected if 𝑟 is more than a threshold
26 Nov 2014 – IGI Journal Club
Evaluation Strategies
1. Synthetic Datasets:
a) CR versus MI values
b) demonstrate outlier rejection
2. Clinical Dataset:
a) Bronze standard: CR versus LMI values
b) Gold standard: landmark correspondences
26 Nov 2014 – IGI Journal Club
• Generate random 1D signals with independent noise – ideal MI or
CR value should be zero
• Calculate image metric (MI or CR)
• CR converges faster to ground truth value
26 Nov 2014 – IGI Journal Club
Outlier rejection
• Traditional outlier detection determine outliers by value
• CR values highly variable according to image deformation,
location
𝐼𝑓 𝐼𝑚 𝜂 𝑟
26 Nov 2014 – IGI Journal Club
Discussion of Synthetic Results
• Demonstrate both MI and CR have the expected results for
random 1D signals
• CR converges to ideal value slightly faster than MI
• CR values hold for a 2D synthetic image
26 Nov 2014 – IGI Journal Club
Dataset Description
• 13 patients
• Preoperative MRI
Acquired ~2 weeks prior to surgery
Gd- enhanced T1W
• Post-resection US
2D US with optical tracking (TA003 tracker,
Polaris optical system)
Freehand movement to acquire 200+ slices
3D pixel-based volumetric reconstruction
26 Nov 2014 – IGI Journal Club
The following parameters tuned using 1 patient dataset
Parameter Effects Best Value
Patch size Locality versus accuracy 73 (343 voxels)
Patch number Accuracy versus computation time
1000
Hierarchical levels Deformation versus smoothness
2
Spacing between B-Spline nodes
Deformation complexity 20 mm
26 Nov 2014 – IGI Journal Club
LMI versus CR
• Selected 1 dataset with a low TME (2.2mm)
• Translated image and calculated LMI, CR
• (0,0) is expected minimum
Making a lot of assumptions (no large deformation, no misalignment)
LMI CR
26 Nov 2014 – IGI Journal Club
Qualitative Assessment of Outlier Detection
26 Nov 2014 – IGI Journal Club
Registration Evaluation
• Expert selected landmarks
Anatomic Structures (sulci bifurcations, vessels, etc)
6/13 cases had landmarks selected twice (1 month apart)
26 Nov 2014 – IGI Journal Club
26 Nov 2014 – IGI Journal Club
26 Nov 2014 – IGI Journal Club
Discussion
• Outlier rejection to ignore
shadow/resected regions
• Comparison to LMI and RaPTOR very
limited
• CR registration driven by matching
strong edges
Matching tumor/resection margin
Difficult to assess if resection misses
tumor edge
26 Nov 2014 – IGI Journal Club
Concluding Remarks
• RaPTOR generally improves MRI-US alignment (2.9 from initial
alignment of 5.9)
• RaPTOR is very quick (30 seconds to register)
• Strong edges drive registration (especially tumor
boundary/resection margin)
• Limited comparisons with other methods (synthetic calculations)
• Dataset available online (BITE Dataset)